An official website of the United States government
This paper analyzes survey nonresponse via the use of multistate Markov models. Typically, such models are used to describe a process or the evolution of a system over time. Using the response history of survey units (SU) initiated into the U.S. Producer Price Index (PPI) over a 10-year period, establishment respondents are shown to move in and out of various, mutually exclusive "states" of response and nonresponse. Using this paradata, a transition probability matrix is constructed and is then used to estimate transition probabilities at various points in a finite time horizon consistent with a seven-year sample. A subset of the paradata are then fed into a Random Forest model with specialized hyperparameters to predict the return of nonrespondents. The probability estimates from the Random Forest can then be used to determine which nonrespondents should be contacted with the goal of mitigating nonresponse overall.